Cybercrime is taking a sharp turn toward automation. A powerful AI-driven tool called ARTEX—originally built for legitimate penetration testing—is now being repurposed by malicious actors to steal vast troves of credentials and digital certificates with minimal human direction. Security researchers found that ARTEX agents have harvested over 190,000 gaming credentials and more than 300 Apple code-signing certificates, though only three certificates were still usable.
ARTEX was once open source, designed to help cybersecurity teams test their defenses. But after it became involved in a breach affecting South Korean banks, its developer made the project closed-source on October 8, 2026. Despite that, criminals continue using forks and backups of the original code, many of which are still uploaded and in circulation.
How the ARTEX Campaigns Operate
Researchers tracking ARTEX activity discovered more than 2,500 systems that host—or once hosted—ARTEX control panels, and over 1,600 of those are still exposed to the internet. Out of ten servers improperly configured, full internal databases were left vulnerable, giving insight into how attackers prompt agents and what models they use.
In one campaign targeting Apple certificate signing services, operators ran ARTEX consoles between October 14 and 29. They focused on ten shops that sell .p12 certificates—files used to sideload iOS apps outside official channels. Using AI models like Deepseek v4.1 Flash and Kimi K3.1 Pro, they processed roughly 36 billion token requests. The total cost for this activity is estimated between $700 and $1,500. Attackers exploited weak security issues—hardcoded credentials in JavaScript, insecure APIs, weak admin passwords, and flaws allowing remote code execution—to reach the certificate libraries. After decrypting keys, checking their validity and revocation status, the ARTEX agents found only three certificates that could still sign apps without restriction.
A separate and larger-scale gaming credential theft operation focused on Chinese mobile game accounts. In just five days, ARTEX agents harvested more than 190,000 login credentials from backend servers for titles like Honor of Kings. About 3.3%—roughly 6,000 accounts—were still active when tested. That campaign cost between $180 and $360 and required almost no operator involvement beyond initial prompting.
Proof That AI-Automated Attacks Are Scaling Fast
Not all ARTEX-driven campaigns succeed. One attempted theft of cryptocurrency from nearly 300 payment domains ended without any funds obtained. Some agents declared success based on generic “success” responses from APIs that hadn’t truly granted access. In another instance, an agent accessed an exposed Amazon S3 storage bucket belonging to an enterprise app service, giving attackers write access—and a potential avenue to swap clean app code for malicious malware—before the owners restored the contents.
Chinese AI models dominate these attacks, especially Deepseek Flash and its variants. When more restricted models (like Opus 5.5) trigger safety guardrails, operators just switch to fewer-constrained, permissive models. Tactics also include using residential proxies to evade firewalls and IP bans. Most active ARTEX servers are in the U.S. (54%), followed by Hong Kong and China. Other AI frameworks—CyberStrike, Cairn, PentAGI—are displaying similar behavior and capabilities.
Key tools being abused include exposed or embedded credentials, unsecured APIs, loose filesystem permissions, and server misconfigurations. Recent advisories warn that secrets and keys hard-coded in code or sitting in public-facing systems headlined these intrusion vectors.
Indicators of compromise linked to ARTEX include specific version tags, IP blocks, and ASNs, particularly in the U.S., Thailand, and occasional nodes in China. Security teams are advised to rotate exposed secrets, eliminate hardcoded credentials, secure APIs with logins, test internet-facing surfaces, and flag unusual spikes in system activity.
“What This Means” Analysis: ARTEX operations mark a turning point in how criminal actors leverage AI: minimal oversight, rapid escalation, and high impact. As AI agents become more accessible, the defenders’ window to react is narrowing. Organizations can no longer wring their hands over policy—they need aggressive posture: continuous secret discovery, infrastructure lockdowns, and AI-aware incident detection. Expect further attacks where speed and automation beat perimeter defenses. The next frontier is whether AI tools will start generating novel exploits rather than just orchestrating human-found vulnerabilities.
